I would like to get started with CUDA development. I plan to get a Tesla C1060 and plug it into a workstation. The applications are all scientific, mostly lots of FFT’s and Linear Algebra.
The workstation mostly has to get data from the disk, load it into the Tesla, and retrieve it. To begin, I am not concerned much with overall speed, I just want to start with a minimal sort of rig, write the software, and measure where the problems are, if any. Then I will know better what to do to speed things up with a more expensive rig if necessary. Does this seem like a reasonable approach?
I’m thinking about a Dell T7400 because it is on the list of qualified workstations, and with a 2.0 Ghz quad core Xeon is inexpensive. I plan to get an FX570 for graphics but have been told that this will still leave space for the Tesla.
My specific questions are:
Do I need to get two Xeon’s in order to really exercise CUDA? Does CUDA even know what to do with multiple processors? Or could I hook it up so that I could take advantage of all 8 cores (two processors in each quad core).
Has anyone used a T7400 with a Tesla, and if so, how did you like it? Did you have any trouble when you plugged the Tesla in and configured all the drivers?
The alternative that I have looked at all involve high priced gaming mother boards which seem overkill to me for an investigative starter system.
Any comments as to whether the T7400 is a suitable approach for a starter system would be welcome.
“2) Has anyone used a T7400 with a Tesla, and if so, how did you like it? Did you have any trouble when you plugged the Tesla in and configured all the drivers?”
A T7400 is sitting in my cube right now as my primary development box. Works absolutely fine with a 2.8GHz quad core Xeon and 8GB of RAM. You don’t need two Xeons to push the card, especially if you’re doing FFTs and linear algebra.
My code can easily max out a GTX 280 with an old single core netburst architecture Xeon… So it doesn’t matter how fast your CPU is if you do everything on the GPU :)
One CPU core per GPU in a multi-GPU system is generally easiest to work with.
Ok, that is a very reasonable approach… so then you buy a “workstation” and a tesla and a quadro?
You don’t need to buy products just because they say “for proffecionals” on the box. Get a normal PC and a GTX260. You’re correct that you don’t need powerful hardware to develop, and please don’t buy into the “profecionals buy products that say ‘for profecionals’” marketing dung. (Buy a Tesla only if you need its memory. Buy an FX570 only if you’re doing CAD. Buy a workstation only if you must have ECC and RAID, for whatever reason.)
P.S. High-end gaming motherboards (already overpriced) do not cost more than workstation ones.
I’ve also been using a Dell T7400 for the last six months doing Cuda development on Linux (RHEL 4, 64-bit). I’ve used both a C870 and a GTX280 with no problems. With pinned memory it delivers around 6 GB/s in either direction to a PCIe v2.0 card.
The only drawback is the power cables. It’s got a 1000W power supply, but only two 6-pin PCIe power connectors. I’m using a 6-pin to 8-pin PCIe power adapter
for my GTX280.